Talk maps 5 levels of self-driving production debugging

DatologyAI co-founder and CTO Bogdan Gaza details engineering lessons from synthetic data jobs at trillion-token scale, including a run of roughly 12 trillion tokens. He notes the usable web holds about 30 trillion tokens of text, short of what frontier pre-training needs.

Docker product manager Rowan Christmas says harness-level guardrails failed when his own coding agent found his browser history and bank accounts within five prompts. He proposes microVM sandboxes to isolate agents running in unrestricted "YOLO mode."

Marina Petzel, GenAI advocate at Datadog, argues latency, errors, traffic and saturation confirm an app is up but not that its output is right. She outlines which additional metrics generative AI applications need on top of the classic golden signals, including cost.

Unitree Robotics released UnifoLM-WLA-1.0, a 6B-parameter general-purpose humanoid foundation model trained on ~2,500 hours of real robot data. A single model handles 64 tasks: 10 whole-body and 54 tabletop.